J. Mech. Phys.

Solids 174 (2023) 105268

Contents lists available at ScienceDirect

Journal of the Mechanics and Physics of Solids

journal homepage: www.elsevier.com/locate/jmps

Computation of effective elastic moduli of rocks using hierarchical

homogenization
Rasool Ahmad a ,∗, Mingliang Liu b , Michael Ortiz a,c , Tapan Mukerji b , Wei Cai a
a Department of Mechanical Engineering, Stanford University, CA 94305, Stanford, USA
b Department of Energy Resources Engineering, Stanford University, CA 94305, Stanford, USA
c
Division of Engineering and Applied Science, California Institute of Technology, CA 91125, Pasadena, USA

ARTICLE INFO ABSTRACT

Keywords: This work focuses on computing the homogenized elastic properties of rocks from 3D micro-
Digital rock physics computed-tomography (micro-CT) scanned images. The accurate computation of homogenized
Elastic moduli properties of rocks, archetypal random media, requires both resolution of intricate underlying
Homogenization
microstructure and large field of view, resulting in huge micro-CT images. Homogenization
FFT solvers
entails solving the local elasticity problem computationally which can be prohibitively expensive
Renormalization
for a huge image. To mitigate this problem, we use a renormalization method inspired scheme,
the hierarchical homogenization method, where a large image is partitioned into smaller
subimages. The individual subimages are separately homogenized using periodic boundary
conditions, and then assembled into a much smaller intermediate image. The intermediate
image is again homogenized, subject to the periodic boundary condition, to find the final
homogenized elastic constant of the original image. An FFT-based elasticity solver is used to
solve the associated periodic elasticity problem. The error in the homogenized elastic constant
is empirically shown to follow a power law scaling with exponent −1 with respect to the
subimage size across all five microstructures of rocks. We further show that the inclusion of
surrounding materials during the homogenization of the small subimages reduces error in the
final homogenized elastic moduli while still respecting the power law with the exponent of
−1. This power law scaling is then exploited to determine a better approximation of the large
heterogeneous microstructures based on Richardson extrapolation.

1. Introduction

Reliable prediction of properties of rocks from the pore-scale micro-computed-tomography (micro-CT) images is the defining
aim of digital rock physics (DRP) (Fredrich et al., 1993; Arns et al., 2001; Dvorkin et al., 2011; Andrä et al., 2013a,b; Saxena et al.,
2017, 2019; Sun et al., 2021). The workflow of DRP starts with the acquisition of pore-scale rock 3D images from micro-CT scans.
The scanned images are next subjected to image processing tools to segment them into underlying constituent materials and pores.
The relevant physical simulations are then performed to compute the target physical properties of interest such as elastic constants,
electrical conductivity, permeability, etc. This work is concerned with computing the effective elastic properties (bulk and shear
moduli) from a given segmented micro-CT image of a rock sample.
A rock is an archetypal example of random heterogeneous solids. The existing theoretical tools to deal with the homogenization
of heterogeneous materials provide rigorous bounds on the homogenized elastic constants. The well known examples of such bounds

∗ Corresponding author.
E-mail addresses: rasool@stanford.edu (R. Ahmad), mliu9@stanford.edu (M. Liu), ortiz@aero.caltech.edu (M. Ortiz), mukerji@stanford.edu (T. Mukerji),
caiwei@stanford.edu (W. Cai).

https://doi.org/10.1016/j.jmps.2023.105268
Received 16 August 2022; Received in revised form 24 January 2023; Accepted 2 March 2023
Available online 8 March 2023
0022-5096/© 2023 Elsevier Ltd. All rights reserved.
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

include the Voigt–Reuss (Voigt, 1910; Reuss, 1929) and Hashin–Shtrikman bounds (Hashin and Shtrikman, 1963). However, these
bounds are generally quite loose and obtained by assuming certain idealized distribution of the phases constituting the heterogeneous
materials. Thus, these rigorous theoretical bounds often fail to provide accurate estimates of microstructure-specific elastic constants
of a heterogeneous material. For a comprehensive review of the homogenization of the elastic constants, readers are referred to
Nemat-Nasser et al. (2013), Zaoui (2002), Charalambakis (2010).
Computational homogenization techniques (Yvonnet, 2019; Khdir et al., 2013; Takano et al., 2000; Vel et al., 2016) are often
used to obtain the microstructure-specific effective elastic constants of materials. The partial differential equation (PDE) of elasticity
(originated from equilibrium, compatibility conditions and material’s constitutive relations) is solved numerically (e.g. using finite
element method or fast fourier transform solvers) to find the stress and strain fields under a prescribed load. The relationship
between the averaged stress and strain values provides information on the effective elastic constants. The elasticity problem in rock
images are generally solved using three numerical methods: (1) finite element method (Garboczi and Berryman, 2001; Roberts and
Garboczi, 2002) (2) Fast Fourier based elasticity solvers (Moulinec and Suquet, 1998; Kabel and Andrä, 2013; Schneider, 2021),
and (3) wave propagation method (Saenger et al., 2000; Saenger, 2008). The modeling of the effect of grain–grain contact on elastic
properties of rocks has been focus of recent computational studies (Sun et al., 2021; Chen et al., 2022). Furthermore, advances in
deep learning, especially computer vision, have been exploited to accelerate the prediction of homogenized elastic constants of rock
images (Karimpouli and Tahmasebi, 2019; Eidel, 2023; Lee et al., 2022).
A major issue in carrying out the computational homogenization arises from the sheer size of the rock sample images that is
required for determining the well-converged effective elastic properties. Rock microstructures contain intricate multiscale features,
i.e. the spatial distribution of minerals and pores that span over a range of length scales. The accuracy of the computationally
homogenized properties depends on both the resolution and the field of view of the digital image (Saxena et al., 2019). A high
resolution is needed to obtain a faithful representation of the fine microstructure, while a large field of view is necessary to ensure
the sample is statistically representative. Unfortunately, refining the resolution of the scanned images while maintaining a large
field of view would result in a very large the number of voxels in the digital image of the rock sample. The computational cost to
perform physical simulations on these huge images quickly becomes prohibitive in terms of time and memory requirements and
undermines the fundamental premises of DRP.
To alleviate the constraints of the available computational resources, here we apply a hierarchical homogenization procedure to
compute the homogenized elastic constants that obviate the need of solving the elasticity equations on a given huge image by brute
force. The basic idea is inspired by the renormalization method (Kadanoff, 1966; Wilson, 1971, 1975, 1983; Fisher, 1998; Efrati
et al., 2014), where we progressively scale down the size of the problem domain by successive applications of a coarse-graining rule
at each scale. In the hierarchical homogenization scheme, a large rock image is partitioned into smaller subimages of size 𝑛 × 𝑛 × 𝑛.
Each smaller subimage is then homogenized computationally and replaced by an equivalent voxel with effective elastic constants.
These voxels are subsequently assembled to get a coarse-grained image with fewer degrees of freedom than the original image we
started with. The assembled coarse-gained image is again homogenized to find the overall homogenized elastic constants of the
sample.
The idea of applying renormalization-based approach to obtain homogenized properties of rocks has attracted some attention in
the past (Kim and Russel, 1985; King, 1989, 1996; Banerjee and Adams, 2004; Green and Paterson, 2007; Karim and Krabbenhoft,
2010; Hanasoge et al., 2017; Hansen et al., 1997; Wei et al., 2019). However, a systematic analysis of the error incurred by the
renormalization-based schemes is largely unavailable. The characterization and control of the associated error constitutes an essential
step in order to confidently apply any approximation method. The present work attempts to characterize the error associated with
the hierarchical homogenization method as applied to computing homogenized elastic moduli of rocks.
The renormalization idea of dividing and reassembling a domain is potentially powerful in reducing the computational
expenses associated with determining the effective elastic properties of heterogeneous materials. However, to our best knowledge,
surprisingly, there does not seem to be a substantial body of research devoted to this area. Banerjee and Adams (2004) employed
the renormalization approach – which they termed recursive cell method – to determine the homogenized elastic properties of
polymer bonded explosive with the aim of reducing computational cost. Their work was limited to relatively small two-dimensional
problems; and employed artificial samples, consisting of random distribution of varying size of circular-shaped hard phases, instead
of real microstructures of materials. They further assumed that the individual subdomains, after coarse graining, can be described
by an isotropic elastic medium. This assumption may be violated for small enough subdomains and add to coarse-graining error.
Moreover, while noticing that the error introduced by the renormalization decreases with increasing size of the subdomains, the
nature of the error has not been examined systematically. The present work seeks to relax some assumptions in the previous studies
and analyze the error caused by the hierarchical homogenization approach in three-dimensional microstructures of real rocks.
In this work, we apply the hierarchical homogenization method to two-phase heterogeneous microstructures obtained from
micro-CT scanning of five different sandstone rock samples. The 3D image of size 𝑁 × 𝑁 × 𝑁 is partitioned into subimages of size
𝑛×𝑛×𝑛 (𝑛 < 𝑁) that are sufficiently larger than the correlation length of the microstructure. Each subimage is then homogenized into
a fully anisotropic elastic medium, represented by a voxel in the coarse-grained image. The error of the hierarchical homogenization
approach is obtained by comparing its prediction of the effective elastic moduli with the brute-force solution of the elasticity PDE
on the original image. We show that the error in the hierarchically homogenized elastic moduli follows a power law scaling with
exponent −1 with respect to 𝑛. The systematic error in the hierarchically homogenized elastic moduli is shown to reduce still
further by introducing a padding of surrounding materials during the homogenization of smaller subimages, and the (𝑛−1 ) scaling
is preserved even with padding. This power law scaling of error is then employed to further reduce the error in the predicted
homogenized elastic constants by Richardson extrapolation. Thus, the main contributions of the present work are

2
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. 1. (a) Segmented image of rock sample B1. Red color denotes quartz mineral (phase 1) and white color represents pore (phase 2). (b) Two point correlation
in sample B1 along the 𝑥-direction. Blue circles are the computed correlation function values, and orange curve is a fit to function exp(−𝑙∕𝜉), where 𝜉 is the
correlation length. 𝜉 is expressed in the units of number of voxels whose physical length is ∼ 2 μm.

Table 1
Statistical properties of the rock samples, including porosity and correlation lengths 𝜉 in 𝑥, 𝑦,
and 𝑧 directions.
Rock sample Porosity 𝜉𝑥 𝜉𝑦 𝜉𝑧
B1 16.51% 13.1 12.8 13.4
B2 19.63% 14.2 15.3 14.5
CG 22.20% 10.7 8.5 11.8
FB1 9.25% 14.4 16.1 15.6
FB2 3.45% 15.9 15.1 15.1

1. A systematic investigation of the error caused by the hierarchical homogenization method as a function of the small subimage
size 𝑛 for 3D real rock images.
2. A demonstration of the existence of a power law with an exponent of −1 between the error and the subimage size 𝑛; and
subsequent utilization of this power law to extrapolate the prediction of homogenized elastic constants in the limit of 𝑛 → ∞.
3. A demonstration of a significant error reduction by use of padding (inclusion of surrounding materials around the subimages)
in the hierarchical homogenization method.

The remaining part of this manuscript is organized as follows. Section 2 briefly discusses the statistical properties of the
rock samples studied in this work. Section 3 presents the hierarchical homogenization method and discusses its main advantages
in reducing the computational costs while controlling the error bounds. Section 4 presents the detailed steps involved in the
homogenization of a single subimage. Section 5 presents the results obtained from the application of the hierarchical homogenization
method to compute the homogenized elastic constants, along with the associated error, of the five digital samples. Section 6
summarizes the main conclusions of this work.

2. Microstructure of rock samples

The rock samples used in this work are the same as those in a previous study (Saxena et al., 2019). These include five
microstructures obtained from three different sandstones: two (B1 and B2) are from the Berea Formation, subangular to subrounded
Mississippian-age sandstone; two (FB1 and FB2) from the Fontainebleau Formation, subrounded to rounded Oligocene age sandstone;
and one (CG) from the Castlegate Formation, subangular to subrounded Mesozoic sandstone. All of the rocks were imaged with a
micro-CT scanner at the image resolution of approximately 2 μm (the physical size of each voxel). The X-ray diffraction (XRD)
analysis demonstrates that all samples mainly consist of quartz mineral along with trace amounts of feldspar, calcite and clay. For
simplicity, we assume single mineralogy where all minerals were treated as quartz. The samples are then segmented into two phases
of quartz (phase 1) and pore (phase 2) using the Otsu algorithm. Fig. 1 shows the segmented images of sample B1 of size 𝑁 3 with
𝑁 = 900. Red colored voxel corresponds to quartz mineral and white colored voxel to pore. The porosity of the five rock samples
are given in Table 1. To facilitate comparisons across rock samples, we assume that the elastic constants of each phase do not vary
across samples: minerals (phase 1) have been assigned the elastic properties of quartz, bulk modulus of 36 GPa and shear modulus
of 45 GPa; and pores (phase 2) have both bulk and shear moduli equal to 0 GPa (Mavko et al., 2003) (for numerical stability, bulk
and shear moduli of the pore phase are taken to be 10−5 GPa). For more information on acquisition and segmentation of micro-CT
imaged rock samples, readers are directed to Saxena et al. (2019).

3
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. 2. A schematic illustration of the two-step hierarchical homogenization scheme. A large image of segmented rock sample of size 𝑁 3 is partitioned into
(𝑁∕𝑛)3 small subimages (shown in red color in the first figure) of size 𝑛3 . Each red subimage is homogenized by solving the local elasticity problem and
assembled into an intermediate image of reduced size (𝑁∕𝑛)3 . The assembled intermediate image is finally homogenized to obtain the homogenized properties
of the original image. Black and white color in the left panel denotes the two individual phases of the rock. The gray level in the middle panel represents the
homogenized value of the relevant properties of the corresponding subimage of size 𝑛3 .

The correlation length is an important characteristic that measures the intrinsic length scale of a microstructure. To compute the
correlation length, we first define a material function 𝑀(𝑖, 𝑗, 𝑘) at every voxel with indices 𝑖, 𝑗, 𝑘 from 1 to 𝑁, such that
{
1 if voxel (𝑖, 𝑗, 𝑘) is occupied by mineral (phase 1)
𝑀(𝑖, 𝑗, 𝑘) = (1)
0 if voxel (𝑖, 𝑗, 𝑘) is occupied by pore (phase 2).
The correlation between two points separated by 𝑙 voxels in the 𝑥-direction (corresponding to 𝑖 index) is given by
⟨𝑀(𝑖, 𝑗, 𝑘)𝑀(𝑖 + 𝑙, 𝑗, 𝑘)⟩ − ⟨𝑀(𝑖, 𝑗, 𝑘)⟩2
Corr𝑥 (𝑙) = , (2)
⟨𝑀 2 (𝑖, 𝑗, 𝑘)⟩ − ⟨𝑀(𝑖, 𝑗, 𝑘)⟩2
where ⟨⋅⟩ represents the average operation over all possible values of indices 𝑖, 𝑗, and 𝑘. We can similarly define correlation functions
in 𝑦- and 𝑧-directions.
Fig. 1(b) shows the two-point correlation function as a function of voxel separation 𝑙 in the 𝑥-direction for the rock sample
B1. The blue circles are the values of the correlation function obtained using Eq. (2), and the orange curves are the best fit of the
exponential function exp (−𝑙∕𝜉), where 𝜉 is the correlation length. For sample B1 the correlation length in 𝑥-direction is 𝜉 = 13.1. The
values of the correlation lengths 𝜉 for the five rock samples are given in Table 1. The correlation lengths in all three directions are
found to be similar, indicating that all the microstructures are isotropic to a good approximation. We observe that the correlation
becomes very small (< 0.05) at distances beyond 3𝜉 (i.e. about 40 for rock sample B1). This information is helpful in deciding the
representative volume element (RVE), i.e. the size of the subimages, for the hierarchical homogenization procedure.

3. Hierarchical homogenization

This section lays out the hierarchical homogenization approach to determine the effective elastic constants of rocks. The goal
of homogenization is to approximate a given random heterogeneous material by a homogeneous material which exhibits the same
elastic response as the original heterogeneous material under some specific loading conditions. In other words, we seek a coarse-
grained description of the heterogeneous material. Instead of homogenizing the whole domain (micro-CT scan image) at once by
solving the elasticity PDE, we follow the renormalization based approach where the problem domain is scaled down progressively
by successive coarse-graining steps.
As schematically depicted in Fig. 2, we start with a large 3D image of size 𝑁 × 𝑁 × 𝑁. This image is then partitioned into (𝑁∕𝑛)3
smaller subimages of size 𝑛 which are shown in the red color in the first subfigure of Fig. 2. Each subimage consists of two phases that
are assumed to be isotropic elastic having different elastic constants. We then seek to replace this small subimage by a single voxel,
i.e. 𝑛3 voxels are decimated into just one voxel that summarizes the (anisotropic) elastic properties of the corresponding subimage.
To accomplish this decimation, we solve the local elasticity problem for all (𝑁∕𝑛)3 smaller subimages separately under periodic
boundary conditions and determine their corresponding average elastic constants, as described in Section 4. In the next step, the
average elastic constants of the subimages are used to assemble a coarse-grained version of the original image. The coarse-grained
image now has (𝑁∕𝑛)3 voxels (see middle panel of Fig. 2) compared with the 𝑁 3 voxels of the original image. The assembled
coarse-grained image retains the low-frequency information of the original image, while the local high-frequency components are
captured in the properties of the small subimages. In the renormalization literature, this coarse-graining/homogenization/decimation
procedure is often repeated multiple times until a fixed point is achieved (Wilson, 1983). However, in this work, we carry out just
two steps of coarse-graining for image size of 900 × 900 × 900, as depicted in Fig. 2. The assembled coarse-grained image is

4
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. 3. Log–log plot of (a) time to solve a single subimage as a function of subimage size, (b) total time elapsed during the hierarchical homogenization (HH)
of a big image of size 𝑁 = 900 as a function of subimage size 𝑛. The calculations are performed using an FFT-based elasticity solver. Blue dots denote the
measured computational time, and the orange line is the fit of the data to a power law.

homogenized one more time to find the effective average elastic constants of the rock sample. We may need to apply more than
two levels of hierarchical homogenization if the original image is very large, e.g. 106 × 106 × 106 .
We now discuss the numerical advantages associated with the hierarchical homogenization scheme. Fig. 3(a) shows the empirical
results of time elapsed for solving an elasticity problem on an image with 𝑛3 voxels. Blue dots are the measured computation time
data, and the orange line is obtained by fitting the data to a power law. For this particular case, the computation time scales as
( )
 𝑛3.66 which is super-linear with respect to the number of voxels (𝑛3 ) in the image. We now consider a large image containing
𝑁 3 voxels. Instead of solving the elasticity PDE on this image, we divide it into (𝑁∕𝑛)3 smaller subimages of size 𝑛3 . Then the time
𝑡 spent in solving (𝑁∕𝑛)3 subimages scales as
( )3 ( ) ( )
𝑁
𝑡∼  𝑛3.66 ∼  𝑛0.66 . (3)
𝑛
Neglecting the time it takes to solve the elasticity problem on the coarse-grained image, the total computation time to solve a
big image with 𝑁 3 voxels thorough hierarchical homogenization (essentially solving (𝑁∕𝑛)3 subimages) scales with the size of the
subimage as (𝑛0.66 ) as shown in Fig. 3(b). The smaller the subimage, the shorter the computation time. The computation time shown
in Fig. 3 and exponent 3.66 are for a specific set of elastic constants of the two phases, and the exact computation time and the
exponent will depend on the contrast between the properties of the underlying constituent phases and the implementation details
of the algorithm. The gain in computation time, nevertheless, will result for any computational algorithm with a time complexity
that is super linear with the number of voxels. We further note that aside from the gain in computation time, the hierarchical
homogenization scheme substantially reduces the memory requirement that would be prohibitive for the full scale simulation,
since the hierarchical homogenization method does not require the simultaneous access to the information about the large image.
The calculations in this section are performed on single Intel(R) Xenon(R) Silver 4216 CPU. For example, the computational
resources available to us limit the largest image that can be solved by brute-force to 10003 voxels due to memory constraints. This
computational efficiency of the hierarchical homogenization method, however, would also introduce approximation error whose
characterization and control is the main focus of this work.

4. Homogenization of a single subimage

This section discusses the two steps involved in determining the average elastic constants of the subimages: (1) extracting the
subimage from the large image; (2) averaging the equilibrium stress and strain fields inside the subimage.
Fig. 4 illustrates the extraction step for a single subimage occupying the domain 𝛺R . In order to compute the effective elastic
constants of the region 𝛺R , it is necessary to solve the elasticity PDE on this domain subjected to appropriate boundary conditions.
For example, it is common practice to apply periodic boundary conditions (PBC) when an FFT-based solver is used. However, the

5
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. 4. Schematic diagram illustrating the extraction and homogenization of the subimages. The goal is to find the average elastic constants of the subimage
shown in the color red occupying the region 𝛺R . In order to mitigate the boundary effects and mimic the original environment experienced by the red subimage
in the original large image, we add extra padding of the surrounding materials to form the blue region 𝛺B ⊇ 𝛺R . We then solve the local elasticity problem in
the extracted blue image 𝛺B subject to periodic boundary conditions. We finally average stress and strain fields inside the red region 𝛺R to find the average
anisotropic elastic constant of the red subimage.

application of PBC (or any other types of boundary conditions) to region 𝛺R introduces an error because the voxels in 𝛺R are
subjected to a different environment compared to that in the original image. Because such an error is introduced at the boundary
of every subimage, we expect the impact of this error on the final homogenized elastic constants to decrease with the size of the
subimage dimension 𝑛.
To control this error due to the subimage boundary conditions, we choose to extract a larger domain of voxels, 𝛺B , which
includes the domain 𝛺R and a padding region, as shown in Fig. 4. The inclusion of this padding region enables, to an extent, the
voxels in 𝛺R to experience a similar environment to that in the original image. The force and moment associated with the boundary
condition are transferred to the 𝛺R region through the surrounding materials. Thus, the distribution of the force and moment will
change at the boundary of 𝛺R , though the net force and moment will remain the same. According to Saint-Venant’s principle, given
a sufficiently thick padding region, the details of the boundary conditions will not matter to the resulting elastic field in domain
𝛺R , as long as the net force and moment transmitted through each boundary face stays the same. Therefore, we expect the error
due to subimage boundary conditions to decay rapidly with the padding region thickness.
In this work, we solve the elasticity PDE on domain 𝛺B subject to PBC. The resulting stress and strain fields are then used to
compute the average stress and strain values in the domain of interest 𝛺R , and its effective elastic constants. Thicker padding region
would lead to more accurate results, but at the expense of higher computational cost. Hence, the choice of the padding thickness is
the result of a trade-off between accuracy and computational cost.
We now solve the elasticity PDE in the domain 𝛺B subjected to PBC. ( The elastic)response of the material at a point 𝒙 ∈ 𝛺B is
( )
prescribed by a position dependent isotropic elastic tensor C𝑖𝑗𝑘𝑙 (𝒙) = 𝐾(𝒙) − 23 𝜇(𝒙) 𝛿𝑖𝑗 𝛿𝑘𝑙 + 𝜇(𝒙) 𝛿𝑖𝑘 𝛿𝑗𝑙 + 𝛿𝑖𝑙 𝛿𝑗𝑘 , where 𝐾(𝒙) and
𝜇(𝒙) are bulk and shear moduli at point 𝒙. Thus, the local stress 𝜎𝑖𝑗 (𝒙) relates to the local strain 𝜀𝑖𝑗 (𝒙) via
( )
1
𝜎𝑖𝑗 (𝒙) = C𝑖𝑗𝑘𝑙 (𝒙)𝜀𝑘𝑙 (𝒙) = 𝐾(𝒙)𝜀𝑘𝑘 (𝒙)𝛿𝑖𝑗 + 2𝜇(𝒙) 𝜀𝑖𝑗 (𝒙) − 𝜀𝑘𝑘 (𝒙)𝛿𝑖𝑗 , (4)
3
where Einstein’s notation is assumed, in which repeated indices are summed over from 1 to 3. We now assume that the domain 𝛺B
is subjected to an average strain 𝐸𝑖𝑗 so that the local strain field can be decomposed into two parts, i.e.

𝜀𝑖𝑗 (𝒙) = 𝐸𝑖𝑗 + 𝜀∗𝑖𝑗 (𝒙), 𝜀∗𝑖𝑗 (𝒙) = 0, (5)

∫𝛺B
where 𝐸𝑖𝑗 corresponds to a homogenized applied strain, and 𝜀∗𝑖𝑗 (𝒙) is the 𝛺B -periodic fluctuation due to the heterogeneity of the
microstructure. The fluctuation in the strain field must satisfy both the compatibility and equilibrium conditions, i.e.
1 ∗
𝜀∗𝑖𝑗 (𝒙) = (𝑢 (𝒙) + 𝑢∗𝑗,𝑖 (𝒙)) ∀𝒙 ∈ 𝛺B , (compatibility condition)
2 𝑖,𝑗 (6)
𝜎𝑖𝑗,𝑗 (𝒙) = 0 ∀𝒙 ∈ 𝛺B , (equilibrium condition)
where 𝑢∗𝑖 (𝒙) is an 𝛺B -periodic displacement field, and ,𝑗 ≡ 𝜕∕𝜕𝑥𝑗 .
Eq. (6) can be solved by using various numerical methods, such as finite element or fast fourier transform (FFT) based solvers, to
obtain equilibrium stress 𝜎𝑖𝑗 (𝒙) and strain 𝜀𝑖𝑗 (𝒙) fields in region 𝛺B . Here we employ an FFT-based method whose basic formulation
is presented in Appendix A. To compute the homogenized elastic constants in domain 𝛺R (excluding the padding), we take the
average of the obtained stress and strain fields over this domain.
1
⟨𝜎⟩R,𝑖𝑗 = 𝑑𝒙 𝜎𝑖𝑗 (𝒙),
|𝛺R | ∫𝛺R
( ) (7)
1
⟨𝜀⟩R,𝑖𝑗 = 𝑑𝒙 𝐸𝑖𝑗 + 𝜀∗𝑖𝑗 (𝒙) .
|𝛺R | 𝛺R
∫

6
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

By definition, the homogenized elastic constant Chomo

R,𝑖𝑗𝑘𝑙
of the subimage satisfies the following condition

⟨𝜎⟩R,𝑖𝑗 = Chomo
R,𝑖𝑗𝑘𝑙 ⟨𝜀⟩R,𝑘𝑙 . (8)

To extract the homogenized elastic constants, it is more convenient to use the Voigt notation, in which the stress and strain
tensors are represented as 6 × 1 column vectors (⟨𝜎⟩ and ⟨𝜀⟩ ) and the elastic stiffness tensor as a 6 × 6 matrix (𝐶 homo ). In the
R R R
Voigt notation, the homogenized stress–strain relation in domain 𝛺R is,

⟨𝜎⟩ = 𝐶 homo ⟨𝜀⟩ . (9)

R R R

In order to compute homogenized stiffness matrix 𝐶 homo , six independent macroscopic strain fields (𝐸𝑖𝑗 ) are imposed and the
R
elasticity PDE is then solved numerically to give six pairs of ⟨𝜎⟩ and ⟨𝜀⟩ vectors. We concatenate the six stress vectors to form
R R
a 6 × 6 matrix 𝛴 , and concatenate the six strain vectors to form a 6 × 6 matrix  . These matrices still satisfy the stress–strain
R R
relation,

𝛴 = 𝐶 homo  . (10)
R R R
Therefore, the homogenized elastic stiffness tensor can be obtained by

𝐶 homo = 𝛴  −1 . (11)
R R R

In general, the resulting homogenized stiffness matrix 𝐶 homo is fully anisotropic. Hence, in the second step of the hierarchical
R
homogenization approach, the coarse-grained image consists of voxels each representing an anisotropic elastic medium, even
if the voxels in the original image represents isotropic elastic medium. The computed voxel stiffness matrix 𝐶 homo is very
R
close
( to being symmetric, e.g., the off-diagonal terms opposite to each differ by ∼ 0.01%. We symmetrize the stiffness matrix
( )T )
𝐶 homo + 𝐶 homo ∕2 to make sure that the voxel stiffness matrices are symmetric.
R R
In this work we use ElastoDict (Anon, 0000; Kabel and Andrä, 2013; Kabel et al., 2016; Schneider et al., 2016) module of
GeoDict software to solve the isotropic elastic problem for small subimages; and to solve fully anisotropic elasticity problem in the
intermediate assembled image, we modify the python code, freely available on webpage (de Geus and Vondřejc, 0000), of which
basic algorithms are detailed in de Geus et al. (2017), Zeman et al. (2017). GeoDict calculations are performed on 16 CPUs and
the python code is run on single CPU of Intel(R) Xenon(R) Silver 4216 machine.
A few remarks on the use of the two elasticity solvers are in order. Geodict solves the elasticity problem using staggered grid
discretization (Schneider et al., 2016) which is stable for porous materials and, thus, suitable for solving the subimages. Currently,
GeoDict cannot handle more then 16 different material types which renders it unsuitable to solve the coarse-grained assembled image
obtained after one step of homogenization (see Fig. 2). The materials of every voxel in the intermediate image can potentially be
different from every other voxel with their own unique elastic properties. The python code (de Geus and Vondřejc, 0000) provides
us with the flexibility of adapting it to solve the intermediate image with each voxel having its own unique anisotropic elastic
constants. The python code is based on uniform grid discretization and is known to be unstable for porous materials (Zeman et al.,
2017; Schneider et al., 2016). However, the coarse-grained assembled image is not porous and the phase contrast in the assembled
image is substantially reduced. We, therefore, do not observe numerical instability while applying uniform grid discretization, as
implemented in the python code, to solve the assembled image. The python code further suffers from a phenomenon associated
with the odd and even number of voxels in the image (de Geus et al., 2017) which, however, does not dilute the main finding of
the work.

5. Results

In this section, we discuss results obtained from the hierarchical homogenization scheme of five different rocks of size 𝑁 3
with 𝑁 = 900. We pay special attention on the influence of subimage size and the padding thickness on the error in the overall
homogenized elastic properties. The large image is partitioned into subimages of size 𝑛3 with 𝑛 = 75, 90, 100, 150, 180, 225, 300.
Padding thicknesses of 0, 10, and 20 are used for each partition scheme.
We first discuss the results for zero padding. For comparison purposes, we also compute the simple arithmetic averages of the
elastic moduli (shear and bulk moduli) of all the subimages of a given partition of the original image which is precisely the Voigt
averaging. The error of each homogenization scheme is computed by comparing with the direct solution of the elasticity PDE
on the original image of size 𝑁 3 . Fig. 5 presents the error in the homogenized elastic moduli as obtained from the hierarchical
homogenization scheme and from simple arithmetic averages as a function of 𝑛 for all five rock samples. The modulus values are
also listed in Table 2. We note that the errors across all cases are positive, indicating a systematic error caused by the periodic
boundary condition used for the homogenization of the subimages (Huet, 1990; Hazanov and Huet, 1994; Kanit et al., 2003).
Fig. 5 shows that the error in the homogenized elastic moduli decreases with increase in subimage size 𝑛 for both hierarchical
homogenization and arithmetic mean values. The monotonic decrease in the error for the hierarchical homogenization indicates that
the rock microstructure is fairly homogeneous at the length scale considered here (𝑛 ≥ 75). This is consistent with the correlation
lengths given in Table 1. The error for the arithmetic mean is larger than the hierarchical homogenization value for the same
subimage size. Therefore, assembling the homogenized subimages into voxels in a coarse-grained image provides a more accurate
description of the overall elastic property than simple averages of the elastic stiffness of the subimages.

7
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. 5. Effect of subimage size 𝑛 on the hierarchically homogenized elastic constants for zero padding case for the five rock samples of size 𝑁 = 900: (a) B1, (b)
B2, (c) CG, (d) FB1, and (e) FB2. In each subfigure, corresponding to a specific rock sample, the variation of errors in the homogenized elastic moduli obtained
from the hierarchical homogenization are plotted on log–log scale as a function of the subimage size 𝑛. The error is calculated by subtracting the corresponding
homogenized elastic moduli of the big image of size 𝑁 = 900 that is obtained by solving the local elasticity problem in the big image. The first panel in each
subfigure shows the result for the bulk modulus and the second panel for the shear modulus. The orange plots show the arithmetic mean (AM) of the elastic
moduli of the small images composing the big image; the blue plots show the result obtained from the hierarchical homogenization (HH); and the green line is
the fit of the hierarchical homogenization values to the function 𝑀∞ + 𝐶𝑛−1 , with 𝑀∞ and 𝐶 being the two fitting parameters.

Table 2
Homogenized elastic moduli (bulk and shear moduli) of the five rock images of size 𝑁 = 900 as a function of small
image size 𝑛 into which the big image is partitioned. The values are obtained by the hierarchical homogenization (HH)
method and by taking arithmetic mean (AM) of the homogenized elastic moduli of the small images.
𝑛
75 90 100 150 180 225 30
Rock type
HH 22.06 21.95 21.78 21.69 21.56 21.54 21.46
Bulk mod (GPa)
AM 22.64 22.38 22.23 21.84 21.70 21.60 21.52
B1
HH 24.51 24.38 24.13 24.05 23.89 23.86 23.73
Shear mod (GPa)
AM 25.59 25.18 24.94 24.34 24.13 23.98 23.85
HH 19.83 19.62 19.42 19.31 19.16 19.15 19.00
Bulk mod (GPa)
AM 20.70 20.25 20.05 19.49 19.32 19.20 19.03
B2
HH 21.58 21.39 21.09 21.01 20.81 20.80 20.58
Shear mod (GPa)
AM 23.10 22.47 22.17 21.32 21.07 20.90 20.63
HH 18.39 18.26 18.11 18.05 17.97 17.93 17.88
Bulk mod (GPa)
AM 18.92 18.62 18.47 18.15 18.06 17.97 17.91
CG
HH 19.78 19.63 19.42 19.37 19.25 19.20 19.13
Shear mod (GPa)
AM 20.65 20.22 20.00 19.54 19.40 19.27 19.18
HH 28.16 28.08 27.99 27.87 27.79 27.77 27.72
Bulk mod (GPa)
AM 28.58 28.36 28.26 27.95 27.86 27.79 27.73
FB1
HH 33.12 33.06 32.89 32.78 32.65 33.61 32.53
Shear mod (GPa)
AM 33.99 33.62 33.44 32.94 32.79 32.66 32.56
HH 33.09 33.06 32.99 32.98 32.95 32.93 32.90
Bulk mod (GPa)
AM 33.31 33.21 33.15 33.02 32.98 32.94 32.91
FB2
HH 40.46 40.42 40.30 40.32 40.27 40.24 40.19
Shear mod (GPa)
AM 40.95 40.76 40.67 40.41 40.34 40.27 40.21

For the hierarchical homogenization, we now quantify the convergence rate of the error with respect to the size of the small
image 𝑛. As shown in Fig. 5, errors in both elastic bulk and shear moduli for the hierarchical homogenization case (blue dots)
decrease with increasing 𝑛 across all five rock samples. Furthermore, on the log–log scale, the variation in error is almost linear,
suggesting a power-law relationship between the error and 𝑛. The fact that error is found by subtracting a fixed value (elastic moduli
of image of size 𝑁 = 900) from the hierarchically homogenized elastic moduli means that the hierarchically homogenized elastic
moduli itself can be well-fitted to a power-law with exponent −1 with an additional additive parameter 𝑀∞ , i.e.

𝑀 HH (𝑛) = 𝑀∞ + 𝐶 𝑛−1 , (12)

where 𝑀 HH (𝑛) stands for both bulk and shear moduli computed from the hierarchical homogenization method using subimages
of size 𝑛3 ; and 𝑀∞ and 𝐶 are the two fitting parameters that, respectively, represent the asymptotic modulus value and the rate
of convergence. The reasoning behind fixing the exponent value to −1 is that the major source of error arises from the interface

8
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Table 3
Homogenized properties of the five different rock samples. 𝑀900 denotes the elastic modulus obtained by directly solving the big
image of size 𝑁 = 900. 𝑀∞ is the asymptotic value of the corresponding elastic modulus that is obtained by fitting the hierarchical
homogenization result to the power law given in Eq. (12). The value of 𝑀∞ is a good approximation for the homogenized elastic
modulus of a big rock sample.
Bulk modulus (GPa) Shear modulus (GPa)
Rock type Porosity (%)
𝑀900 𝑀∞ 𝐶 𝑀900 𝑀∞ 𝐶
B1 16.51 21.30 21.26 58.95 23.51 23.50 74.89
B2 19.63 18.82 18.77 76.44 20.30 20.32 91.28
CG 22.20 17.77 17.71 48.38 18.96 18.91 61.60
FB1 9.25 27.63 27.57 45.27 32.37 32.33 59.65
FB2 3.45 32.85 32.85 18.13 40.10 40.13 23.53

between the two neighboring small subimages. The homogenized subimages capture the bulk microstructure features present within
the image. However, an error is introduced by the hierarchical homogenization method at the interface between subimages due to
the application of periodic boundary conditions. Therefore, we expect the error in the final elastic moduli to scale as the ratio
between the surface area and volume of the subimages which is 𝑛−1 .
In Fig. 5, the results of the fitting procedure are plotted in green. The qualities of the power-law scaling with fix exponent value
of −1 are found to be good for both elastic moduli across all five rock samples. The values of these fitting parameters, 𝑀∞ and 𝐶,
along with the elastic moduli 𝑀900 of the image size 900 for all five rock samples are given in Table 3. The values of 𝑀∞ represents
the asymptotic value of the corresponding elastic moduli, and approximates very well the homogenized elastic moduli of a large
image. Thus, the hierarchical homogenization scheme can be employed to compute the homogenized elastic moduli of a big rock
image without ever solving the local elasticity problem in the big original image, thus gaining in time and avoiding the memory
constraints encountered routinely in the application of computational methods to large domains.
We further observe that the error in the arithmetic mean (AM) elastic moduli follows a power law with exponent −1.5. This
scaling law for the AM moduli has been established theoretically in Gloria and Otto (2011, 2012), Gloria et al. (2015). Although
the scaling is better (−1.5 versus −1), the error in the AM modulus for a given subimage size is strictly greater than that in the
corresponding hierarchically homogenized modulus.
Fig. 6 demonstrates the effect of including padding materials surrounding the subimages on the hierarchically homogenized
elastic constants (bulk and shear moduli) of the corresponding big image. Three padding thickness, 0, 10, 20, are used, and the
resulting error are plotted in blue, orange and green color in Fig. 6. The error in the hierarchically homogenized elastic moduli of
the big image consistently decreases with increasing padding thickness across all the five rock samples. This means that the same
error can be achieved using smaller subimages when some padding materials are included. For instance, for the bulk modulus of B1
rock sample, the error incurred by subimage size of 300 with zero padding is about the same as that obtained from subimage size
of 175 with padding thickness of 10. As discussed in Section 3 and shown in Fig. 3, the overall computational time will be shorter
for hierarchical homogenization using subimages of size 175 with padding thickness of 10 than that for subimages of size 300 with
zero padding.
The hierarchically homogenized elastic constants are further observed to follow the same power law 𝑀∞ + 𝐶𝑛−1 , for all three
padding thickness values. The curves obtained from fitting the power law to the simulation data are plotted in solid lines in Fig. 6.
Moreover, the fitting parameters (𝑀∞ and 𝐶) are presented in Table 4 of the five rock samples, each with three padding thickness
values. The parameter 𝑀∞ − representing the asymptotic hierarchically homogenized elastic constants − is almost unchanged with
padding thickness for both bulk and shear modulus. The value of 𝐶, however, generally decreases with increasing padding thickness
which indicates a faster convergence to the asymptotic value. There is a relatively little change in the value of 𝐶 for FB2 rock sample.
This is presumably caused by the dominance of one phase (phase 1) in the composition of the material.

6. Discussion and conclusion

In this work, we apply a renormalization inspired method, the hierarchical homogenization method, to determine the average
elastic constants of a random heterogeneous materials. The microstructures obtained from micro-CT scan of five real sandstone
rock samples – two from Berea Formation, one from Castlegate Formation, two from Fontainebleau Formation – are used as the
heterogeneous materials. The micro-CT scanned three-dimensional images are segmented into two phases (pore and mineral) are
assigned to two different isotropic elastic materials.
In the hierarchical homogenization scheme, a large image is partitioned into multiple disjoint smaller subimages. The subimages
are homogenized separately by solving local elasticity problems. The homogenized subimages are then assembled and homogenized
to calculate the average elastic constants of the original big image. The hierarchical homogenization thus obviates the need to solve
the local elasticity problem in the big domain and, thus, alleviating the prohibitive requirement of computational resources in terms
of computational time and memory. This is particularly advantageous for cases where capturing the small length-scale features of
the microstructure is important to produce accurate elastic properties. For instance, recent work on digital rock physics (Saxena
et al., 2019) indicates that the insufficient resolution of the micro-CT scans leads to a systematic error in the computed average
elastic constants. Application of high-resolution imaging modality would dramatically increase the number of voxels for a given

9
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. 6. Effect of the padding thickness on the hierarchically homogenized elastic constants for different padding thickness for the five rock samples of size
𝑁 = 900: (a) B1, (b) B2, (c) CG, (d) FB1, and (e) FB2. In each subfigure, corresponding to a specific rock sample, the variation of errors in the homogenized
elastic moduli obtained from the hierarchical homogenization are plotted on log–log scale as a function of the subimage size 𝑛 for the three values of padding
thickness: 0, 10, 20. The error is calculated by subtracting the corresponding homogenized elastic moduli of the big image of size 𝑁 = 900 that is obtained by
solving the local elasticity problem in the big image. The first panel in each subfigure shows the result for the bulk modulus and the second panel for the shear
modulus. The three colors correspond to the three values of the padding thickness: blue for 0, orange for 10, and green for 20. The dashed lines with circle
markers show the simulation result and the solid curve is the fit of the hierarchical homogenization values to the function 𝑀∞ + 𝐶𝑛−1 , with 𝑀∞ and 𝐶 being
the two fitting parameters. For (c) CG rock microstructure, the errors in bulk modulus (for padding thickness 20 and 𝑛 = 100, 180) and shear modulus (padding
thickness 20 and 𝑛 = 100) become negative, and, thus, corresponding absolute values are plotted with thick cross. Across all the five rocks, the error decreases
with increasing padding thickness for a fixed subimage size.

Fig. 7. Distribution of bulk modulus in the assembled rock images after first step of hierarchical homogenization using different sizes 𝑛 of smaller images. A
random two-dimensional slice of the three-dimensional image is plotted for each case. The rows correspond to different rock samples and the column are for
different sizes 𝑛 of the small image used in the hierarchical homogenization. The first column (𝑛 = 1) presents the bulk modulus distribution in the original
image without any homogenization. We observe that increasing the size of the smaller image increases the homogeneity of the assembled image.

10
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Table 4
The result of the fitting of the power law expressions to the hierarchically homogenized elastic moduli (bulk
and shear moduli) of the five rock samples with varying padding thicknesses. 𝑀∞ is the asymptotic value of the
corresponding elastic modulus and 𝐶 controls the convergence rate of the power law.
Rock type Padding thickness Bulk Modulus (GPa) Shear Modulus (GPa)
𝑀∞ 𝐶 𝑀∞ 𝐶
0 21.26 58.95 23.50 74.89
B1 10 21.30 25.98 23.50 35.03
20 21.30 11.94 23.50 17.59
0 18.77 76.44 20.32 91.28
B2 10 18.77 37.49 20.26 50.29
20 18.79 16.74 20.27 26.15
0 17.71 48.38 18.91 61.60
CG 10 17.74 15.64 18.92 25.42
20 17.77 0.84 18.94 7.12
0 27.56 45.27 32.33 59.65
FB1 10 27.58 31.21 32.31 45.57
20 27.59 19.73 32.31 31.59
0 32.85 18.13 40.13 23.53
FB2 10 32.84 15.50 40.10 22.92
20 32.84 11.67 40.09 19.42

representative volume and, therefore, render solving the local elasticity problem in the full image impractical. The hierarchical
homogenization approach makes it feasible to predict the effective elastic moduli from these large images.
The main finding of this work is the 𝑛−1 scaling of the systematic error in the final homogenized elastic constants with the
dimension 𝑛 of the subimages. The error is caused mainly by two following reasons. The first is the complete neglect of the effect
of surrounding materials on the homogenized elastic constants of the smaller subimages. The second is the mismatch between
the boundary conditions used for the subimages during the first homogenization step and in the original image subject to its
own loading. The homogenization of the subimages is performed under periodic boundary condition, however, the subimages are
generally under mixed boundary conditions when they are part of the original image. Thus, the error arises from the inaccurate
description of materials near the surface of the smaller subimages. This rationalizes the observed exponent of −1 associated with
the power-law scaling of the error in hierarchical homogenization. This power law scaling, furthermore, enables a good estimate
of the average elastic constants of the original big image. We have verified that all these properties continue to hold for a different
set of elastic properties of the two phases of the microstructure as shown in Appendix B. We, however, do not have a theory that
conclusively explains the empirical observation of the power law. We hope that this work will stimulate future, more rigorous
theoretical investigation into the origin of this scaling behavior of error associated with the hierarchical homogenization.
The understanding of the cause of the systematic error in hierarchical homogenization also suggests a way to reduce this error.
To incorporate the effect of the surrounding material and mimic the boundary condition experienced by the subimage as part of
the big image, we introduce a padding of surrounding materials around the subimages during its homogenization. The error in the
hierarchically homogenized elastic moduli progressively reduces with increase in the thickness of padding materials. The power
law scaling of the hierarchically homogenized elastic moduli, with the exponent of −1 and almost the same asymptotic value, is
preserved across all the values of the padding thickness. The convergence of the hierarchically homogenized elastic moduli to the
asymptotic value with respect to the subimage size accelerates with increase in the padding thickness. This further bolsters the
robustness of the use of the power-law to extract the unbiased estimate of the elastic moduli of a big image using the hierarchical
homogenization method.
The computational homogenization of the coarse-grained assembled image is performed by solving the local elasticity problem
numerically using FFT-based elasticity solver on uniform grid (de Geus et al., 2017). The FFT elasticity solver using uniform grid
discretization satisfies equilibrium and compatibility conditions only for odd-sized grids, i.e., images with odd number of voxels. If
the number of grid points is even, both conditions cannot be satisfied simultaneously due to the treatment of the Nyquist frequency
component (Vondřejc et al., 2015; de Geus et al., 2017). This difference in the odd- and even-sized (size of the intermediate
assembled images, i.e. 𝑁∕𝑛, varies as 12, 10, 9, 6, 5, 4, 3) is reflected in the slightly jagged pattern of the blue line in Figs. 5 and
6. We, however, note that this particular phenomenon does not dilute the main result of the present work which are the power law
scaling of the error and the beneficial effect of inclusion of surrounding padding materials during homogenization of the subimages.
The size of the subimages to partition the original big image is chosen so that each subimage is sufficiently larger than the
correlation length of the underlying microstructure. Thus, the elastic constants of the voxels of the assembled image are distributed
about a mean value. The variance of the elastic constants decreases with increase in the subimage size rendering assembled image
increasingly homogeneous, as shown in Fig. 7. This fact is also the basis of the trend of increasing error with decreasing subimage
size. However, this trend does not hold as we continue to decrease the size of subimage. As illustrated in Fig. 7, once the subimage
size is comparable to the correlation length of the microstructure, the assembled image begins to retain the microstructural details
and resemble the heterogeneous original image. Thus, we expect that below a cutoff size, the trend in error vs subimage size reverses,
resulting in decrease in error with decrease in subimage size. This reverse trend has indeed been observed.

11
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

The hierarchical homogenization method along with error behavior, as presented in this work, can be used to homogenize other
properties of rocks other than elastic constants. In fact, the hierarchical homogenization method has been applied to obtain effective
permeability of rocks without investigating associated error. It would be interesting to extend this method for heat conductivity,
electrical conductivity and magnetic properties of rocks. Again, error characterization and control would be essential to apply the
hierarchical homogenization method to these areas.
In conclusion, this work shows that the hierarchical homogenization method in combination with the power law scaling of the
associated error can be employed to homogenize a huge domain of random heterogeneous materials such as rocks. This will enable
the efficient computation of effective properties of very high-resolution images of the underlying heterogeneous materials.

CRediT authorship contribution statement

Rasool Ahmad: Conception and design of study, Methodology, Analysis and interpretation, Writing of the original manuscript,
Revision of the manuscript. Mingliang Liu: Analysis and interpretation, Writing of the original manuscript. Michael Ortiz:
Analysis and interpretation, Writing of the original manuscript. Tapan Mukerji: Analysis and interpretation, Writing of the original
manuscript. Wei Cai: Conception and design of study, Methodology, Analysis and interpretation, Writing of the original manuscript,
Revision of the manuscript.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgments

We acknowledge Shell for financial support and for providing the digital rock images. We thank Dr. Nishank Saxena at Shell
for his constant support of the project. The authors also would like to thank Math2Market for providing the GeoDict software at a
discount, and providing technical support.

Appendix A. FFT-based solver for elasticity problem

The most important component of the computational homogenization is solving the equilibrium and compatibility equations on
a given domain (6). Since micro-CT scan and subsequent segmentation provides information about the rock microstructure in a
three-dimensional grid, use of FFT-based solver is well-suited for solving the local elasticity problem subjected to periodic boundary
conditions. The FFT-based solver was introduced by Moulinec and Suquet (1994) and further expanded and advanced by several
works (Moulinec and Suquet, 1998; Kabel et al., 2014; Vondřejc et al., 2014, 2015; de Geus et al., 2017; Zeman et al., 2017; Leute
et al., 2022). This section describes the basic formulation of the FFT-based solver, borrowing from Zeman et al. (2017) which closely
follows the standard finite element method (FEM) approach to obtain the discretized form of Eqs. (6). For a more comprehensive
review of FFT-based homogenization, readers are referred to Schneider (2021), Lucarini et al. (2022)
The discretization of local problem begins with reformulating the local problem, Eq. (6), into the weak form, i.e.

𝑑𝒙 𝛿𝜀∗𝑖𝑗 (𝒙) 𝜎𝑖𝑗 (𝒙, 𝐸𝑖𝑗 + 𝜀∗𝑖𝑗 (𝒙)) = 0, (A.1)

∫𝛺
where 𝛿𝜀∗𝑖𝑗 (𝒙) is a test strain field which itself must satisfy the compatibility and periodic boundary conditions.
The FEM approach satisfies the compatibility requirement and periodic boundary condition of the test strain field by expressing
it in terms of a periodic test displacement field, 𝛿𝜀∗𝑖𝑗 (𝒙) = (𝛿𝑢∗𝑖,𝑗 (𝒙) + 𝛿𝑢∗𝑗,𝑖 (𝒙))∕2. The FFT-based approach, on the other hand, works
directly with strain field without expressing it in terms of the underlying displacement field. In this case the compatibility of the test
strain field is enforced via a four-rank self adjoint projection operator G𝑖𝑗𝑘𝑙 (𝒙) which project any symmetric two-rank non-compatible
tensor field 𝜁𝑖𝑗 (𝒙) to a compatible part, i.e.

𝛿𝜀∗𝑖𝑗 (𝒙) = [G ⋆ 𝜻]𝑖𝑗 (𝒙) = 𝑑𝒚 G𝑖𝑗𝑙𝑘 (𝒙 − 𝒚) 𝜁𝑘𝑙 (𝒚) (A.2)

∫𝛺
is a compatible test strain field obtained by projecting non-compatible symmetric test field 𝜁𝑖𝑗 (𝒙). In the above expression, ⋆ denotes
convolution operation. The projection operator has a block diagonal form in the Fourier space, and is given by
̂ 𝑖𝑗𝑙𝑚 (𝒌) = 𝛿𝑖𝑚 𝑔̂𝑗𝑙 (𝒌), where
G
⎧
⎪0 if 𝒌 = 𝟎 (A.3)
𝑔̂𝑗𝑙 (𝒌) = ⎨ 𝑘𝑗 𝑘𝑙
⎪ ‖𝒌‖2 otherwise,
⎩

12
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. A.8. A schematic illustration of the discretization of the image and the coordinate system used in the FFT elasticity solver in two dimensions. The coordinate
system is placed at the center of the image discretized in 𝑁 number of pixels in each direction. The center of each pixel is the discretization point. A discretization
point is characterized by two integers 𝑝1 and 𝑝2 , where −𝑁∕2 < 𝑝1 , 𝑝2 < 𝑁∕2.

where the hat symbol ̂. over a quantity denotes the fourier transform of that quantity; 𝒌 is a vector in the reciprocal space. The
convolution operation is then performed efficiently in the Fourier space as
[ ]
[G ⋆ 𝜻]𝑖𝑗 (𝒙) =  −1 Ĝ 𝑖𝑗𝑘𝑙 (𝒌) 𝜁̂𝑙𝑘 (𝒌) (𝒙), (A.4)
where  −1 [⋅] is inverse Fourier transform operation.
Thus, the weak form can be written as

𝑑𝒙 [G ⋆ 𝜻]𝑖𝑗 (𝒙) 𝜎𝑖𝑗 (𝒙, 𝐸𝑖𝑗 + 𝜀∗𝑖𝑗 (𝒙)) = 𝑑𝒙 𝜁𝑖𝑗 (𝒙) [G ⋆ 𝝈]𝑖𝑗 (𝒙, 𝐸𝑖𝑗 + 𝜀∗𝑖𝑗 (𝒙)) = 0, (A.5)
∫𝛺 ∫𝛺
where the self-adjoint property of projection operator G𝑖𝑗𝑘𝑙 (𝒙) is exploited to obtain the second expression of Eq. (A.5) which has
the advantage that the test field 𝜁𝑖𝑗 (𝒙) now does not need to satisfy the compatibility condition and can be a member of the set of
symmetric and periodic rank-two tensor fields.
Eq. (A.5) is now poised for discretization in the vein of the standard Galerkin method. To illustrate the discretization procedure,
here we consider a two-dimensional problem. As shown in Fig. A.8, the information about the microstructure is provided in the
form of a square grid of size 𝑁 × 𝑁, i.e. 𝑁 nodes along each coordinate axis. Microstructure data obtained from micro-CT scan has
this underlying grid structure (three-dimensional), and therefore belongs to the category under consideration. We further suppose
that the number of pixels along each dimension 𝑁 is odd; this will simplify the exposition of the procedure by focusing on the
essential features of FFT-based methods. The case of even number of pixels 𝑁 will be discussed later.
The coordinate of a grid point is identified by a pair of numbers 𝒑 = [𝑝1 , 𝑝2 ] as
𝒑
𝒙𝑁 = 𝑝1 𝒆1 + 𝑝2 𝒆2 , (A.6)
where index 𝒑 is sampled from a reduced set of index Z2𝑁 , which is a set of pair of integers, each constrained to lie between −𝑁∕2
and 𝑁∕2, i.e.
{ }
| 𝑁 𝑁 𝑁 𝑁
𝒑 ∈ Z2𝑁 = Z2 ||− ≤ 𝑝1 ≤ ,− ≤ 𝑝1 ≤ . (A.7)
| 2 2 2 2
The next step in the discretization of the weak form, Eq. (A.5), is definition of basis functions. In FFT-based methods, fundamental
trigonometric polynomials are used as the basis functions. The fundamental trigonometric polynomial associated with node 𝒑 is
defined as
( ) ( )
𝒑 1 ∑ 2𝜋𝑖 2𝜋𝑖
𝜑𝑁 (𝒙) = exp − 𝒑 ⋅ 𝒒 exp 𝒒⋅𝒙 ∀𝒒 ∈ Z2𝑁 . (A.8)
𝑁 2 𝑁 𝑁
2 𝒒∈Z𝑁

The fundamental trigonometric polynomials, unlike FEM basis function, have global support, and still possess suitable interpolation
properties, such as
𝒑 ( 𝒒 )
𝜑 𝒙 = 𝛿𝒌𝒎 ,
∑𝑁 𝒑 𝑁
𝜑𝑁 (𝒙) = 1 ∀𝒙 ∈ 𝛺. (A.9)
𝒑∈Z2𝑁

13
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. B.9. Effect of subimage size on the hierarchically homogenized elastic constants for zero padding case for the five rock samples of size 𝑁 = 900: (a) B1,
(b) B2, (c) CG, (d) FB1, and (e) FB2. The orange plots show the arithmetic mean (AM) of the elastic moduli of the small images composing the big image; the
blue plots show the result obtained from the hierarchical homogenization (HH); and the green line is the fit of the hierarchical homogenization values to the
function 𝑀∞ + 𝐶𝑛−1 , with 𝑀∞ and 𝐶 being the two fitting parameters.

The first identity in above equation states that the fundamental trigonometric polynomial takes value 1 at the associated node, and
it assumes value 0 at every other node. The second identity is a statement of the partition of unity.
Discretization of Eq. (A.5) is performed by expressing both test field and the solution fluctuating strain field as linear combinations
of the fundamental trigonometric polynomials
∑ 𝒑 𝒑
𝜁𝑖𝑗 (𝒙) = 𝜑𝑁 (𝒙) 𝜁𝑖𝑗 ,
𝒑∈Z2𝑁
∑ 𝒒 ∗𝒒
(A.10)
𝜀∗𝑖𝑗 (𝒙) = 𝜑𝑁 (𝒙) 𝜀𝑖𝑗 ,
𝒒∈Z2𝑁
𝒑 ∗𝒑
where 𝜁𝑖𝑗 and 𝜀𝑖𝑗 are values of the test and strain field, respectively, at node 𝒑. We then insert Eq. (A.10) into Eq. (A.5) to obtain

∑ ⎛ ∑ 𝒒 ⎞
𝜑𝑁 (𝒙) 𝜁𝑖𝑗 [G ⋆ 𝝈]𝑖𝑗 ⎜𝒙, 𝐸𝑖𝑗 + 𝜑𝑁 (𝒙) 𝜀𝑖𝑗 ⎟ ≈ 0,
𝒑 𝒑 ∗𝒒
𝑑𝒙
∫𝛺 ⎜ ⎟
𝒑∈Z2𝑁 ⎝ 𝒒∈Z2𝑁 ⎠
∑ ∑ ⎛ ∑ 𝒒 ⎞
𝒑 𝒑 ⎜𝒙𝒎 , 𝐸𝑖𝑗 + ∗𝒒 ⎟ (A.11)
𝜑𝑁 (𝒙𝒎 ) 𝜁 [G ⋆ 𝝈]𝑖𝑗 𝜑 (𝒙 𝒎
) 𝜀 𝑖𝑗 ⎟ ≈ 0,
𝑁 𝑖𝑗 ⎜ 𝑁 𝑁 𝑁
𝒎∈Z2 𝒑∈Z2
𝑁
⎝ 𝒒∈Z2𝑁 ⎠
∑ ( )
∗𝒎
𝜁𝑖𝑗𝒎 [G ⋆ 𝝈]𝑖𝑗 𝒙𝒎 𝑁 , 𝐸 𝑖𝑗 + 𝜀 𝑖𝑗 ≈ 0.
𝒎∈Z2

The second line of the equation is obtained by numerically integrating the first line using the trapezoidal quadrature rule. The third
line is the result of the application of the first property of the fundamental trigonometric polynomials as presented in Eq. (A.9). The
final step in the discretization is application of the fact that the third line of Eq. (A.11) is satisfied for all admissible values of the
test fields 𝜁𝑖𝑗𝒎 . This condition leads to the following set of equations
( )
∗𝒎
[G ⋆ 𝝈]𝑖𝑗 𝒙𝒎 𝑁 , 𝐸𝑖𝑗 + 𝜀𝑖𝑗 ≈0 ∀𝒎 ∈ Z2𝑁 . (A.12)

This equation is further simplified by the use of local constitutive relation for each node. Since material associated with each node
is assumed to be linear isotropic elastic with its own elastic constant tensor, we reach to following equation
( ) ( )
G ⋆ C𝒎 ∶ 𝜺∗𝒎 = −G ⋆ C𝒎 ∶ 𝑬 , (A.13)
where C∗ is elastic constant tensor of the material associated with node 𝑚. This is a linear equation in nodal fluctuation strain 𝜀∗𝑚
𝑖𝑗 ,
and is solved by using conjugate gradient method. The use of conjugate gradient method also ensures compatibility of 𝜀∗𝒎 𝑖𝑗 field at
every step of the iteration. For further implementation details on FFT-based elasticity solver, interested readers are referred to de
Geus et al. (2017), Zeman et al. (2017).
The main advantages of the FFT-based elasticity solvers include: (1) The solver operates directly on the images of microstructure
and obviate the need for generating meshes to represent the microstructure, (2) the global stiffness matrix need not be assembled,
(3) use of efficient fast fourier transform algorithm to perform the convolution operation, and (4) natural integration of periodic
boundary conditions which is desirable for homogenization problems.

14
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Fig. B.10. Effect of the padding thickness on the hierarchically homogenized elastic constants for different padding thickness for the five rock samples of size
𝑁 = 900: (a) B1, (b) B2, (c) CG, (d) FB1, and (e) FB2. The three colors correspond to the three values of the padding thickness: blue for 0, orange for 5, and
green for 10. The dashed lines with circle markers show the simulation result and the solid curve is the fit of the hierarchical homogenization values to the
function 𝑀∞ + 𝐶𝑛−1 , with 𝑀∞ and 𝐶 being the two fitting parameters.

Table B.5
Homogenized elastic moduli (bulk and shear moduli) of the five rock images of size 𝑁 = 900 as a function of small
image size 𝑛 into which the big image is partitioned. The values are obtained by the hierarchical homogenization
method and by taking arithmetic mean of the homogenized elastic moduli of the small images.
𝑛
75 90 100 150 180 225 300
Rock type
HH 61.48 61.33 61.04 60.96 60.79 60.79 60.65
Bulk mod (GPa)
AM 62.93 62.38 62.10 61.33 61.10 60.95 60.81
B1
HH 26.92 26.89 26.83 26.81 26.77 26.76 26.73
Shear mod (GPa)
AM 27.23 27.11 27.05 26.88 26.83 26.79 26.76
HH 56.61 55.31 56.04 55.92 55.71 55.75 55.53
Bulk mod (GPa)
AM 58.58 57.74 57.34 56.29 56.03 55.88 55.60
B2
HH 25.20 24.87 25.07 25.05 24.99 24.99 24.93
Shear mod (GPa)
AM 25.64 25.45 25.37 25.13 25.06 25.02 24.95
HH 52.82 52.65 52.41 52.31 52.18 52.12 52.03
Bulk mod (GPa)
AM 53.88 53.34 53.08 52.50 52.35 52.21 52.10
CG
HH 23.88 23.83 23.77 23.75 23.71 23.69 23.67
Shear mod (GPa)
AM 24.14 24.00 23.94 23.79 23.75 23.71 23.68
HH 75.62 75.49 75.19 75.09 74.92 74.89 74.79
Bulk mod (GPa)
AM 77.09 76.42 76.10 75.36 75.15 74.96 74.84
FB1
HH 31.42 31.41 31.36 31.34 31.31 31.30 31.28
Shear mod (GPa)
AM 31.69 31.57 31.52 31.38 31.34 31.31 31.29
HH 89.12 89.02 88.80 88.82 88.73 88.71 88.64
Bulk mod (GPa)
AM 90.08 89.70 89.52 89.02 88.87 88.75 88.68
FB2
HH 35.25 35.24 35.21 35.21 35.20 35.19 35.18
Shear mod (GPa)
AM 35.40 35.34 35.31 35.24 35.22 35.20 35.19

Appendix B. Microstructure with non-zero elastic stiffness in both phases

In this section, we present the results of hierarchical homogenization where the two phases of the microstructures are assigned
different elastic properties from those used in the main text. We seek to show that the power law scaling with exponent −1, and
padding-effect is not an accident of the particular values of the elastic constants of the two phases (e.g. stiffness of the pore phase
being zero), and persist across different phase properties. In particular, bulk and shear moduli of phase 1 are, respectively, 98.04 GPa
and 37.59 GPa; and bulk and shear moduli of phase 2 are, respectively, 9.8.0 GPa and 3.76 GPa. Thus, both phases have finite
stiffness and phase 1 is ten times stiffer than phase 2.
Fig. B.9 presents the homogenized elastic moduli obtained from the hierarchical homogenization and arithmetic average for
zero padding case. Fig. B.10 shows the effects of padding thickness of the surrounding materials around the subimages on the
hierarchically homogenized elastic moduli. Furthermore, the values obtained from these computations are presented in Table B.5
and B.6. The overall behaviors of these results are the same to that presented in the main text where the two phases have the

15
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Table B.6
Homogenized properties of the five rock samples. 𝑀900 denotes the elastic modulus obtained by directly solving the big image
of size 𝑁 = 900. 𝑀∞ is the asymptotic value of the corresponding elastic modulus that is obtained by fitting the hierarchical
homogenization result to the power law given in Eq. (12). The value of 𝑀∞ is a good approximation for the homogenized elastic
modulus of a big rock sample.
Bulk modulus (GPa) Shear modulus (GPa)
Rock type Porosity (%)
𝑀900 𝑀∞ 𝐶 𝑀900 𝑀∞ 𝐶
B1 16.51 60.47 60.40 78.87 26.67 26.68 18.64
B2 19.63 55.21 55.31 99.77 24.86 24.87 24.18
CG 22.20 51.77 51.87 75.83 23.60 23.61 20.73
FB1 9.25 74.51 74.64 81.89 31.23 31.24 14.22
FB2 3.45 88.50 88.55 43.24 35.15 35.16 55.84

properties of quartz and pore. In particular, we observe that hierarchically homogenized elastic moduli follow the power law with
the exponent −1 across all padding thickness values (0, 5, 10) studied; error in the hierarchically homogenized elastic moduli reduces
with padding thickness for a fixed subimage size 𝑛; convergence of hierarchically homogenized elastic moduli to the corresponding
asymptotic value with subimage size 𝑛 becomes faster with increasing padding thickness. Thus, the emergence of the power law
and effect of padding materials are expected to be robust and independent of the particular values of the phase properties.

References

Andrä, H., Combaret, N., Dvorkin, J., Glatt, E., Han, J., Kabel, M., Keehm, Y., Krzikalla, F., Lee, M., Madonna, C., Marsh, M., Mukerji, T., Saenger, E.H., Sain, R.,
Saxena, N., Ricker, S., Wiegmann, A., Zhan, X., 2013a. Digital rock physics benchmarks-part I: Imaging and segmentation. Comput. Geosci. 50, 25–32.
Andrä, H., Combaret, N., Dvorkin, J., Glatt, E., Han, J., Kabel, M., Keehm, Y., Krzikalla, F., Lee, M., Madonna, C., Marsh, M., Mukerji, T., Saenger, E.H., Sain, R.,
Saxena, N., Ricker, S., Wiegmann, A., Zhan, X., 2013b. Digital rock physics benchmarks-part II: Computing effective properties. Comput. Geosci. 50, 33–43.
Anon, 0000. http://www.geodict.de/modules/dicts/elastodict.php.
Arns, C.H., Knackstedt, M.A., Pinczewski, W.V., Lindquist, W.B., 2001. Accurate estimation of transport properties from microtomographic images. Geophys. Res.
Lett. 28, 3361–3364.
Banerjee, B., Adams, D.O., 2004. On predicting the effective elastic properties of polymer bonded explosives using the recursive cell method. Int. J. Solids Struct.
41, 481–509.
Charalambakis, N., 2010. Homogenization techniques and micromechanics. A survey and perspectives. Appl. Mech. Rev. 63, 1–10.
Chen, B., Xiang, J., Latham, J.P., 2022. Influence of inter-grain cementation stiffness on the effective elastic properties of porous Bentheim sandstone. J. Rock
Mech. Geotech. Eng. https://doi.org/10.1016/j.jrmge.2022.06.009.
de Geus, T.W., Vondřejc, J., 0000. http://goosefft.geus.me/.
de Geus, T.W., Vondřejc, J., Zeman, J., Peerlings, R.H., Geers, M.G., 2017. Finite strain FFT-based non-linear solvers made simple. Comput. Methods Appl. Mech.
Engrg. 318, 412–430.
Dvorkin, J., Derzhi, N., Diaz, E., Fang, Q., 2011. Relevance of computational rock physics. Geophysics 76.
Efrati, E., Wang, Z., Kolan, A., Kadanoff, L.P., 2014. Real-space renormalization in statistical mechanics. Rev. Modern Phys. 86, 647–667.
Eidel, B., 2023. Deep cnns as universal predictors of elasticity tensors in homogenization. Comput. Methods Appl. Mech. Engrg. 403, 115741.
Fisher, M.E., 1998. Renormalization group theory: Its basis and formulation in statistical physics. Rev. Modern Phys. 70, 653–681.
Fredrich, J.T., Greaves, K.H., Martin, J.W., 1993. Pore geometry and transport properties of Fontainebleau sandstone. Int. J. Rock Mech. Min. Sci. 30, 691–697.
Garboczi, E.J., Berryman, J.G., 2001. Elastic moduli of a material containing composite inclusions: effective medium theory and finite element computations.
Mech. Mater. 33, 455–470.
Gloria, A., Neukamm, S., Otto, F., 2015. Quantification of ergodicity in stochastic homogenization : optimal bounds via spectral gap on Glauber dynamics. Invent.
Math. 199, 455–515.
Gloria, A., Otto, F., 2011. An optimal variance estimate in stochastic homogenization of discrete elliptic equations. Ann. Probab. 39, 779–856.
Gloria, A., Otto, F., 2012. An optimal error estimate in stochastic homogenization of discrete elliptic equations 1. Ann. Appl. Probab. 22, 1–28.
Green, C.P., Paterson, L., 2007. Analytical three-dimensional renormalization for calculating effective permeabilities. Transp. Porous Media 68, 237–248.
Hanasoge, S., Agarwal, U., Tandon, K., Koelman, J.M.A., 2017. Renormalization group theory outperforms other approaches in statistical comparison between
upscaling techniques for porous media. Phys. Rev. E 96, 1–10.
Hansen, A., Roux, S., Aharony, A., Feder, J., Jøssang, T., Hardy, H.H., 1997. Real-space renormalization estimates for two-phase flow in porous media. Transp.
Porous Media 29, 247–279.
Hashin, Z., Shtrikman, S., 1963. A variational approach to the theory of the elastic behaviour of multiphase materials. J. Mech. Phys. Solids 11, 127–140.
Hazanov, S., Huet, C., 1994. Order relationships for boundary conditions effect in heterogeneous bodies smaller than the representative volume. J. Mech. Phys.
Solids 42, 1995–2011.
Huet, C., 1990. Application of variational concepts to size effects in elastic heterogeneous bodies. J. Mech. Phys. Solids 38, 813–841.
Kabel, M., Andrä, H., 2013. Fast numerical computation of effective elastic moduli of porous materials. Rep. Fraunhofer ITWM 224, 1–16.
Kabel, M., Böhlke, T., Schneider, M., 2014. Efficient fixed point and Newton–Krylov solvers for FFT-based homogenization of elasticity at large deformations.
Comput. Mech. 54, 1497–1514.
Kabel, M., Fliegener, S., Schneider, M., 2016. Mixed boundary conditions for FFT-based homogenization at finite strains. Comput. Mech. 57, 193–210.
Kadanoff, L.P., 1966. Scaling laws for ising models near tc. Physics 2, 263–272.
Kanit, T., Forest, S., Galliet, I., Mounoury, V., Jeulin, D., 2003. Determination of the size of the representative volume element for random composites: Statistical
and numerical approach. Int. J. Solids Struct. 40, 3647–3679.
Karim, M.R., Krabbenhoft, K., 2010. New renormalization schemes for conductivity upscaling in heterogeneous media. Transp. Porous Media 85, 677–690.
Karimpouli, S., Tahmasebi, P., 2019. Image-based velocity estimation of rock using convolutional neural networks. Neural Netw. 111, 89–97.
Khdir, Y.K., Kanit, T., Zaïri, F., Naït-Abdelaziz, M., 2013. Computational homogenization of elastic–plastic composites. Int. J. Solids Struct. 50, 2829–2835.
Kim, S., Russel, W.B., 1985. Modelling of porous media by renormalization of the stokes equations. J. Fluid Mech. 154, 269–286.
King, P.R., 1989. The use of renormalization for calculating effective permeability. Transp. Porous Media 4, 37–58.
King, P.R., 1996. Upscaling permeability: Error analysis for renormalization. Transp. Porous Media 23, 337–354.

16
R. Ahmad et al. Journal of the Mechanics and Physics of Solids 174 (2023) 105268

Lee, Y., Yim, J., Hong, S., Min, K.B., 2022. Application of artificial neural network for determining elastic constants of a transversely isotropic rock from a
single-orientation core. Int. J. Rock Mech. Min. Sci. 160.
Leute, R.J., Ladecký, M., Falsafi, A., Jödicke, I., Pultarová, I., Zeman, J., Junge, T., Pastewka, L., 2022. Elimination of ringing artifacts by finite-element projection
in FFT-based homogenization. J. Comput. Phys. 453, 110931.
Lucarini, S., Upadhyay, M.V., Segurado, J., 2022. FFT based approaches in micromechanics: fundamentals, methods and applications. Modelling Simul. Mater.
Sci. Eng. 30, 023002.
Mavko, G., Mukerji, T., Dvorkin, J., 2003. The Rock Physics Handbook: Tools for Seismic Analysis of Porous Media. Cambridge University Press.
Moulinec, H., Suquet, P., 1994. A fast numerical method for computing the linear and nonlinear mechanical properties of composites. Mech. Solids.
Moulinec, H., Suquet, P., 1998. A numerical method for computing the overall response of nonlinear composites with complex microstructure. Comput. Methods
Appl. Mech. Engrg. 157, 69–94.
Nemat-Nasser, S., Hori, M., Achenbach, J.D., 2013. Micromechanics: Overall Properties of Heterogeneous Materials. Elsevier Science.
Reuss, A., 1929. Berechnung der fließgrenze von mischkristallen auf grund der plastizitätsbedingung für einkristalle. ZAMM Z. Angew. Math. Mech. 9, 49–58.
Roberts, A.P., Garboczi, E.J., 2002. Elastic properties of model random three-dimensional open-cell solids. J. Mech. Phys. Solids 50, 33–55.
Saenger, E.H., 2008. Numerical methods to determine effective elastic properties. Internat. J. Engrg. Sci. 46, 598–605.
Saenger, E.H., Gold, N., Shapiro, S.A., 2000. Modeling the propagation of elastic waves using a modified finite-difference grid. Wave Motion 31, 77–92.
Saxena, N., Hofmann, R., Alpak, F.O., Dietderich, J., Hunter, S., Day-Stirrat, R.J., 2017. Effect of image segmentation & voxel size on micro-CT computed effective
transport & elastic properties. Mar. Pet. Geol. 86, 972–990.
Saxena, N., Hofmann, R., Hows, A., Saenger, E.H., Duranti, L., Stefani, J., Wiegmann, A., Kerimov, A., Kabel, M., 2019. Rock compressibility from microcomputed
tomography images: Controls on digital rock simulations. Geophysics 84, WA127–WA139.
Schneider, M., 2021. A review of nonlinear FFT-based computational homogenization methods. Acta Mech. 232, 2051–2100.
Schneider, M., Ospald, F., Kabel, M., 2016. Computational homogenization of elasticity on a staggered grid. Internat. J. Numer. Methods Engrg. 105, 693–720.
Sun, Z., Salazar-Tio, R., Duranti, L., Crouse, B., Fager, A., Balasubramanian, G., 2021. Prediction of rock elastic moduli based on a micromechanical finite element
model. Comput. Geotech. 135, 104149.
Takano, N., Zako, M., Ishizono, M., 2000. Multi-scale computational method for elastic bodies with global and local heterogeneity. J. Comput.-Aided Mater. Des.
7, 111–132.
Vel, S.S., Cook, A.C., Johnson, S.E., Gerbi, C., 2016. Computational homogenization and micromechanical analysis of textured polycrystalline materials. Comput.
Methods Appl. Mech. Engrg. 310, 749–779.
Voigt, W., 1910. Lehrbuch der Kristallphysik: (mit Ausschluss der Kristalloptik). B.G. Teubner.
Vondřejc, J., Zeman, J., Marek, I., 2014. An FFT-based galerkin method for homogenization of periodic media. Comput. Math. Appl. 68, 156–173.
Vondřejc, J., Zeman, J., Marek, I., 2015. Guaranteed upper–lower bounds on homogenized properties by FFT-based galerkin method. Comput. Methods Appl.
Mech. Engrg. 297, 258–291.
Wei, S., Shen, J., Yang, W., Li, Z., Di, S., Ma, C., 2019. Application of the renormalization group approach for permeability estimation in digital rocks. J. Pet.
Sci. Eng. 179, 631–644.
Wilson, K.G., 1971. Renormalization group and critical phenomena. I. Renormalization group and the Kadanoff scaling picture. Phys. Rev. B 4, 3174–3183.
Wilson, K.G., 1975. The renormalization group: Critical phenomena and the kondo problem. Rev. Modern Phys. 47, 773–840.
Wilson, K.G., 1983. The renormalization group and critical phenomena. Rev. Modern Phys. 55, 583–600.
Yvonnet, J., 2019. Computational Homogenization of Heterogeneous Materials with Finite Elements. Springer International Publishing.
Zaoui, A., 2002. Continuum micromechanics: Survey. J. Eng. Mech. 128, 808–816.
Zeman, J., de Geus, T.W., Vondřejc, J., Peerlings, R.H., Geers, M.G., 2017. A finite element perspective on nonlinear FFT-based micromechanical simulations.
Internat. J. Numer. Methods Engrg. 111, 903–926.

17